Nov. 19, 2025

Using AI for Real-Time Control of Rapid-Fire Lasers

By Charlie Osolin

The next generation of high-energy laser systems, such as those that would be required in future inertial fusion energy (IFE) power plants, are capable of firing 10 times or more per second, much too fast for human operators to track and adjust their performance on the fly.

Not too fast, though, for artificial intelligence (AI), as Lawrence Livermore National Laboratory (LLNL) researchers and their colleagues are now discovering. A team from LLNL, UCLA, and California State University Channel Islands (CSUCI) recently demonstrated the potential for machine learning and AI to exert instantaneous control of a rapid-fire laser system.

“This demonstration represents a paradigm shift in real-time control and stabilization over high-rep-rate plasma systems,” said Derek Mariscal, group leader for high-repetition-rate (HRR) high energy density science within the NIF & Photon Science Directorate’s Advanced Photon Technologies (APT) program.

“This is a very powerful outcome.”

—LLNL researcher Abhik Sarkar

By integrating AI models into high-power HRR laser experiments, the team precisely manipulated plasma characteristics—such as size and density evolution—without directly adjusting the laser or target. Instead, they trained the models to link system inputs to desired plasma outcomes.

“This breakthrough represents a major step forward for high energy density science and inertial fusion energy,” Mariscal said, “paving the way for next-generation AI-driven experimental platforms.”

“To successfully commercialize fusion energy,” added Tammy Ma, director of LLNL’s Livermore Institute for Fusion Technology, “AI-driven prediction, control, and operations will be needed for a fusion power plant. This work by Derek, Abhik, and the team is not only a highly innovative and significant step forward for IFE, but it also demonstrates LLNL’s leadership in rapid prediction and calculations for HED science and plasma physics.”

Schematic of a potential inertial fusion energy power plant — An inertial fusion energy power plant would use high-power lasers to create continual fusion ignition reactions from a steady stream of pellets containing the hydrogen isotopes deuterium and tritium. The pellets would be fired into the plant at a rate of about 600 per minute. The plant's lasers would precisely converge on each pellet, causing them to ignite and give off 50 to 100 times more energy than went in. That excess energy could then be converted into a clean, abundant source of electricity and connected to the power grid. The target chamber would be surrounded by a lithium blanket that, when irradiated by energetic neutrons, would produce tritium to fuel new targets. Laser fusion power plants could be similar in size to today’s large baseload power plants, and just one could meet the energy needs of a major American city.

Machine learning already has demonstrated the ability to foresee the outcome of National Ignition Facility (NIF) inertial confinement fusion (ICF) experiments to an accuracy of better than 70 percent. And LLNL researchers are also deploying AI agents on two of the world’s most powerful supercomputers to engineer new ICF targets.

“To get the maximum or intended performance in a HRR laser experiment in the presence of changes to the system that you know a human likely couldn’t respond to in time,” Mariscal said, “you need to be able to go fast, and you need to be able to handle large amounts of data.”

Many Benefits

In a project called Data-Driven Atomic Physics (DDAP), the APT-led team combined a machine-learning model with direct AI model control to manipulate the laser in HRR laser plasma experiments at the UCLA Phoenix Laser Laboratory. “That (combination) has all kinds of important benefits,” Mariscal said. “It’s super fast, and it loves large amounts of data.”

The experimental work was carried out through a new Joint Initiative collaborative mode through LaserNetUS whereby users work as a team with facilities, as opposed to solely as users. The LLNL team worked closely with UCLA to stand up a new experimental physics and industrial control system (EPICS) at the Phoenix laboratory “that basically gave us ‘hooks’ to plug in an AI control system,” Mariscal said.

“Now we have these new platforms where we can do experiments much, much faster than we can do on NIF or even mid-scale facilities,” he said. “We can now get many orders of magnitude more data in the same amount of time than we would on one of the other facilities.”

Experiments at the ELI Beamlines Facility in the Czech Republic utilized technology that was previously developed through a Scientific Discovery Through Advanced Computing (SciDAC) project. There, researchers used the LLNL-developed High-Repetition-Rate Advanced Petawatt Laser System (L3-HAPLS) to generate protons in the facility’s ELIMAIA laser-plasma ion accelerator.

Using a machine learning accelerated framework for optimizing the design of multiple simulations, originally developed by LLNL’s Cognitive Simulation team, data generated during individual batches of experiments were analyzed and fed to an optimizer that suggested new laser pulse shapes to human operators. While a successful strategy, system limitations and a lack of integrated infrastructure prohibited real-time operations.

The DDAP team developed a first-of-its-kind method for real-time control over the plasma itself. “Here, instead of a Gaussian (statistical probability distribution) process typically used in smart optimization algorithms, we directly used a deep learning model for control,” Mariscal said, “and that has all kinds of important benefits—it’s very fast and it handles large amounts of data.”

“An AI control system creates an abstraction that lets you largely ignore the complex relationships between lasers and targets, enabling targeted real-time control,” added LLNL researcher Abhik Sarkar, who worked with UCLA and CSUCI to develop the DDAP project’s infrastructure and control system.

Experimental setup for the HRR experiments. A rotating, movable cylindrical target filled with nitrogen gas at different pressures is driven by up to 10,000 variable laser shots, firing at 1 Hz (one shot per second). The resulting plasma expands, creating a “blast wave” captured by a microchannel plate (MCP) camera. The background gas pressure inhibits plasma expansion when observed at a fixed point in time. An AI controller learns the relationship between control points, such as the lens or waveplate position, and the plasma blast wave.

For their studies, the researchers set up a HRR laser experiment with a variety of variable characteristics, such as the energy and intensity of the laser and the pressure of a nitrogen gas-filled chamber—“a lot of different knobs you can turn that affect the physics of the system,” Mariscal said.

After the laser created a plasma on the solid target, the resulting expansion into the ambient gas created a “blast wave,” which was affected by the position of the various “knobs.”

They then trained the machine-learning model with an initial set of data mapping from controls to outputs characterized in terms of physical parameters, such as the location of the blast wave in space and time.

Finally, they defined an outcome—a desired location and timing of the blast wave—and a deep learning model predicted the control parameters that would be needed to reach that outcome. “The idea was that we should be able to make an instantaneous prediction and therefore rapidly change the controls between shots” for direct AI model control, Mariscal said.

AI embedded into a laser system — By embedding artificial intelligence directly into a laser system, LLNL researchers and their colleagues were able to exert real-time control over targeted physics parameters with sub-10 percent accuracy. (Top subplot) White points represent measured blast wave positions while red points show user-defined locations.

To analyze the data in real-time, the team employed specialized diagnostic instruments paired with deep learning algorithms developed under the Laboratory Directed Research and Development Program.

An AI Detective?

“This (demonstration) is a very powerful outcome,” Sarkar said, “and it sets the foundation for building more advanced forms of AI control. “With AI capabilities increasing rapidly, it could potentially act like a detective—figuring out the causes of failures or instabilities and ultimately refining system control and performance.”

Mariscal emphasized the importance of collaboration.

“Having Abhik, the UCLA team, and our partners at CSU Channel Islands managing the data infrastructure and interfaces was crucial,” he said. “That’s what makes this all possible. When we arrived, I didn’t have to touch anything in the lab—it was just a matter of coding. That’s the dream: data flows and controls are easily accessible, so it all comes down to the algorithm.”

Looking Ahead

The success of real-time AI control in high-repetition-rate laser experiments signals a new era for high energy density physics and fusion research. As AI evolves, it will accelerate discoveries, boost precision, and help overcome challenges once thought impossible. These advances bring us closer to a future of clean, abundant fusion energy.

With each breakthrough, the line between human ingenuity and machine intelligence blurs, paving the way for smarter, more autonomous laboratories. The energy solutions of tomorrow, and the next generation of fusion power plants, may be powered by the AI-driven innovations emerging today.

Joining Mariscal and Sarkar on the DDAP team are Blagoje Djordjević, Rich London, Madison Martin, Matt Hill, Raspberry Simpson, Elizabeth Grace, and Ronnie Shepherd from LLNL. The experimental team included (Co-PI) Robert Dorst (now at LLNL), Carmen Constantin, Enrique Cisneros, Jackson Roth, and Chris Niemann from UCLA; and Scott Fiester from CSUCI. The team also thanked everyone who has inspired and supported this work behind the scenes, including IFE and AI initiatives at LLNL.

More Information

“From Ignition to Energy,” Science & Technology Review, July-August 2025

“LLNL Researchers Employed an AI-Driven Model to Predict Fusion Ignition Shot,” NIF & Photon Science News, August 29, 2025

“Predicting fusion ignition at the National Ignition Facility with physics-informed deep learning, ” Science, August 14, 2025

“ELI and LLNL Strengthen Collaboration,” NIF & Photon Science News, July 7, 2025

“LLNL Pushes Frontier of Fusion Target Design with AI,” NIF & Photon Science News, August 11, 2025

“Big Ideas Lab Podcast: How AI Is Reshaping Science at LLNL,” NIF & Photon Science News, July 1, 2025

“Machine Learning Optimizes High-Power Laser Experiments,” NIF & Photon Science News, May 20, 2024

“High Performance Computing, AI, and Cognitive Simulation Helped LLNL Conquer Fusion Ignition,” NIF & Photon Science News, June 21, 2023

Video: “Igniting Scientific Discovery with AI and Supercomputing,” NIF & Photon Science News, June 10, 2024

“Toward machine-learning-assisted PW-class high-repetition-rate experiments with solid targets,” Physics of Plasmas, July 15, 2024

Follow us on X: @lasers_llnl